Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

GLOCALFAIR: Jointly Improving Global and Local Group Fairness in Federated Learning

Syed Irfan Ali Meerza, Luyang Liu, Jiaxin Zhang, Jian Liu

TL;DR

GLOCALFAIR, a client-server co-design fairness framework that can jointly improve global and local group fairness in FL without the need for sensitive statistics about the client's private datasets, is proposed.

Abstract

Federated learning (FL) has emerged as a prospective solution for collaboratively learning a shared model across clients without sacrificing their data privacy. However, the federated learned model tends to be biased against certain demographic groups (e.g., racial and gender groups) due to the inherent FL properties, such as data heterogeneity and party selection. Unlike centralized learning, mitigating bias in FL is particularly challenging as private training datasets and their sensitive attributes are typically not directly accessible. Most prior research in this field only focuses on global fairness while overlooking the local fairness of individual clients. Moreover, existing methods often require sensitive information about the client's local datasets to be shared, which is not desirable. To address these issues, we propose GLOCALFAIR, a client-server co-design fairness framework that can jointly improve global and local group fairness in FL without the need for sensitive statistics about the client's private datasets. Specifically, we utilize constrained optimization to enforce local fairness on the client side and adopt a fairness-aware clustering-based aggregation on the server to further ensure the global model fairness across different sensitive groups while maintaining high utility. Experiments on two image datasets and one tabular dataset with various state-of-the-art fairness baselines show that GLOCALFAIR can achieve enhanced fairness under both global and local data distributions while maintaining a good level of utility and client fairness.

GLOCALFAIR: Jointly Improving Global and Local Group Fairness in Federated Learning

TL;DR

GLOCALFAIR, a client-server co-design fairness framework that can jointly improve global and local group fairness in FL without the need for sensitive statistics about the client's private datasets, is proposed.

Abstract

Federated learning (FL) has emerged as a prospective solution for collaboratively learning a shared model across clients without sacrificing their data privacy. However, the federated learned model tends to be biased against certain demographic groups (e.g., racial and gender groups) due to the inherent FL properties, such as data heterogeneity and party selection. Unlike centralized learning, mitigating bias in FL is particularly challenging as private training datasets and their sensitive attributes are typically not directly accessible. Most prior research in this field only focuses on global fairness while overlooking the local fairness of individual clients. Moreover, existing methods often require sensitive information about the client's local datasets to be shared, which is not desirable. To address these issues, we propose GLOCALFAIR, a client-server co-design fairness framework that can jointly improve global and local group fairness in FL without the need for sensitive statistics about the client's private datasets. Specifically, we utilize constrained optimization to enforce local fairness on the client side and adopt a fairness-aware clustering-based aggregation on the server to further ensure the global model fairness across different sensitive groups while maintaining high utility. Experiments on two image datasets and one tabular dataset with various state-of-the-art fairness baselines show that GLOCALFAIR can achieve enhanced fairness under both global and local data distributions while maintaining a good level of utility and client fairness.
Paper Structure (37 sections, 3 theorems, 27 equations, 8 figures, 7 tables, 1 algorithm)

This paper contains 37 sections, 3 theorems, 27 equations, 8 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Primer on Federated Learning and Fairness
  4. Federated Learning
  5. Fairness Notion
  6. Global and Local Group Fairness
  7. Design of GLocalFair
  8. Design Overview
  9. Local Fairness by Imposing Constraints
  10. Local Fairness Constraints
  11. Constraint Optimization
  12. Global Fairness by Fairness-aware Clustering
  13. Experimental Evaluation
  14. Experimental Setup
  15. Federated Datasets
  16. ...and 22 more sections

Key Result

Theorem 1

Define $\mathcal{M}$ as the set of all stochastic $(m+1)*(m+1)$ matrices, $\Lambda := \Delta^{m+1}$ as the $(m+1)$ dimensional simplex, where $m$ is the number of functional constraints, and assume that each $\hat{g_i}$ upper bounds the corresponding $g_i$ ($g_i$ represents the $i^{th}$ constraint). where $f_L(\theta)$ is the binary cross-entropy loss that we want to minimize. Let denote the objec

Figures (8)

  • Figure 1: An illustrative example of bias in classification when considering two sensitive gender groups (i.e., male vs. female) in FL with three participating clients. At the global scale, the model is fair due to the equal true-positive rate (or false-negative rate). However, when considering the model's fairness on each client, the classification result is biased. Note that the model performs differently between clients mainly due to the client non-IIDness in FL, such as different distributions in racial and age groups.
  • Figure 2: Illustration of the proposed GLocalFair.
  • Figure 3: Lorenz curve for different client model weights.
  • Figure 4: Distribution of EOD and DP-Dis on individual clients for the CelebA datasets (Results on other datasets can be found in Supplementary Material \ref{['subsec:local_fairness_other_datasets']}).
  • Figure 5: Comparison of feature heat maps produced by different baselines and GLocalFair (More results can be found in Supplementary Material \ref{['additional_interpretability']}).
  • ...and 3 more figures

Theorems & Definitions (5)

  • Theorem 1
  • proof
  • Corollary 1
  • Lemma 1
  • proof